AIGC Detector & Rewriter Skill: A Minimal-Edit Approach to Thesis AI-Writing Risk

The problem this skill addresses: rewriting that makes a thesis look more AI-generated

The pitch behind most AI rewriting tools is fluency. They smooth transitions, vary vocabulary and tighten logic. The README argues this is backwards for thesis revision: text that becomes more fluent and more polished can look more AI-generated to a detector, not less. The skill is built for the opposite strategy, keeping as much original author wording as possible and avoiding over-polishing, including preserving non-native but acceptable academic style.

The intended user is narrow. This is a skill for master's thesis text in English, Chinese or a mix of both, quantitative, qualitative or mixed-methods, where the author has already received an external AI detection report from Turnitin AI, CNKI AIGC, GPTZero-style or Originality-style tools. It is not a general-purpose paraphraser and not a writing assistant. The README's stated goal is to preserve the author's thesis while reducing structural AI-writing risk, and it says directly that the goal is not to create perfect native-speaker prose. If that trade-off sounds unappealing, the tool is aimed at someone else.

How the mechanism works: layered risk analysis, language routing and a version ledger

The skill checks AI-risk signals at five levels, according to the README's table: sentence level (repeated openings, uniform rhythm, excessive transitions), paragraph level (mechanical claim-explanation-conclusion structure), section level (repeated literature review, hypothesis or results-reporting templates), chapter level (long continuous detector-highlighted fragments) and document level (repeated architecture across chapters). That layering is the core design decision. A single paragraph can be clean while a whole chapter repeats the same hypothesis-development shape, and the skill is built to see both.

Language routing is the second mechanism. English thesis text goes through a D1 to D17 academic AI-risk dimension framework. Chinese thesis text uses Chinese AI-risk phrase and structure checks. Mixed-language documents are routed segment by segment according to language. Qualitative text adds qualitative-specific risk checks, and quantitative text adds methodology and empirical-reporting structure checks. The README's justification is concrete: this prevents Chinese thesis text from being incorrectly judged by English-only AI-writing indicators, which is a real failure mode when a single model handles both languages.

External detector reports are treated as calibration signals, not as ground truth the model can simulate. The README lists Turnitin AI, CNKI AIGC, GPTZero-style and Originality-style reports as inputs. The repository layout includes a references/ directory and a KNOWN_LIMITATIONS.md file alongside SKILL.md, TEST_CASES.md and MANIFEST.md, so the operational content lives in SKILL.md rather than the README. The README also shows a version ledger among the core capabilities, which matters because the allowed operations are conservative enough that you need to see what changed between passes.

Installing the skill and running a first chapter pass

The README does not give install commands. It identifies the project as an Openclaw and Hermes-Agent skill, and the repository root holds SKILL.md, MANIFEST.md and references/, so the entry point is the skill manifest rather than a package on a registry. Treat this as a skill you load into an Openclaw or Hermes-Agent environment, then invoke against thesis text. If your agent host expects a manifest-driven skill directory, SKILL.md is what you read first; the README is a description of behaviour, not a setup guide.

The practical first use is a single chapter, not the whole thesis. The README describes a chapter-by-chapter revision workflow with external re-testing, and the core principle is sentence-by-sentence and paragraph-by-paragraph editing. So the first pass should be one chapter that an external report already flagged. The README states that DOCX, Markdown and text-based workflows are supported, and that the skill can use external detector reports as calibration signals, so those are the inputs to prepare. What you should see afterwards is a per-passage risk listing plus a change ledger, not a rewritten chapter. The README's forbidden-operations list rules out whole-chapter rewriting, so if a run returns freshly generated prose for the entire chapter, something is wrong with how the skill was invoked, not with the chapter.

Allowed edits versus forbidden operations: where the conservative line sits

The allowed operations are deliberately small: substituting individual words, reordering clauses where safe, splitting overly balanced sentences, recasting repeated sentence openings, reducing formulaic transitions, making existing reasoning more visible, reorganizing paragraph architecture without changing meaning, relocating existing author reasoning when traceable, and preserving or increasing word count when required. Notice that every item is a transformation of text that already exists. Nothing in the list creates new content.

The forbidden list is the more informative half. It bans whole-paragraph regeneration, whole-chapter rewriting, back-translation, detector-bypass rewriting, third-party humanizer rewriting, fabricating data, citations, examples, quotes or researcher voice, changing variables, hypotheses, coefficients, p-values or conclusions, and claiming that an external detector score has improved without a comparable external re-test. That last prohibition is the strictest thing in the README. It means the skill will not tell you your score went down. You have to go get a new report.

This is a genuine limitation, not a marketing hedge. A tool that refuses to predict detector output is less satisfying to use than one that hands you a number. The README's own before-and-after graphic shows an AI feature value dropping from 60% to 18%, but the surrounding text frames detector scores as external evidence, and the forbidden-operations list forbids claiming improvement without a comparable external re-test. Read the graphic as an illustration of the workflow, not as a promise about your chapter.

When this is the wrong tool

Three cases stand out. First, if you need a guaranteed score, this is the wrong project. The README states that it does not guarantee any AI detector score and is not a detector-bypass tool. Any tool that does promise a number is making a claim this one refuses to make.

Second, if your draft is genuinely thin, minimal editing will not fix it. The skill cannot fabricate data, citations, examples, quotes or researcher voice, and it cannot change variables, hypotheses, coefficients, p-values or conclusions. If a chapter lacks authorial interpretation, the skill can only make existing reasoning more visible. It cannot supply reasoning that was never written. For qualitative passages with decorative quotations or weak authorial interpretation, the README lists the pattern as a risk it targets, but targeting a pattern is not the same as generating the missing analysis.

Third, if your document is not an academic thesis, the routing logic is wasted. The D1 to D17 framework, the Chinese AI-risk phrase checks and the methodology and empirical-reporting structure checks are all tuned to thesis conventions: literature reviews, hypothesis development, results reporting. A blog post or a product page will not match those templates, and the layered analysis has nothing to compare against. The project's own scope statement is English and Chinese academic writing, and that boundary is worth respecting.

How it differs from a general-purpose humanizer

A general-purpose humanizer takes a passage and rewrites it in different words, usually with the aim of producing text that reads as if a person wrote it. The transformation is generative: the output is new prose that carries the same meaning. The AIGC Detector & Rewriter Skill does the opposite. Its allowed operations are substitutions, clause reordering, sentence splitting, opening recasts and paragraph reorganisation, all applied to the original text. The README explicitly forbids third-party humanizer rewriting and back-translation, which is a direct statement that this is not that category of tool.

The difference shows up in what each approach does to your citations and numbers. A generative rewrite can silently alter a coefficient or drop a citation marker while producing smoother prose. A minimal-edit pass that is forbidden from touching variables, hypotheses, coefficients, p-values or conclusions cannot do that, at the cost of producing text that still sounds like you wrote it, including the non-native phrasing the README says it preserves on purpose. If you want the smoother output, a humanizer will give it to you. If you want the smoother output and your own data intact, you have to check the result line by line, which is the work this skill is designed to avoid.

Maintenance, licence and the upgrade cost of a skill that lives in SKILL.md

The repository is not archived. The last push was on 2026-06-19, and the most recent release listed is v2.21.0 on 2026-06-16, preceded by v2.19.1 on 2026-06-10. That is a short gap between releases, so the project was moving quickly through June 2026. Whether it still is, the README does not say.

The licence is Apache-2.0, which permits commercial and academic use and requires that you retain the licence and notice files, and that modified files carry prominent change notices. The repository root contains a LICENSE file. Apache-2.0 also includes a patent grant and an explicit disclaimer of warranty. None of that is legal advice; if you are distributing a modified skill, read the LICENSE file.

The upgrade cost is the interesting part. Because the operational content sits in SKILL.md with a MANIFEST.md and a references/ directory, and because the README advertises a version ledger, your change history is part of the artefact. Upgrading means checking whether the D1 to D17 framework or the Chinese phrase checks changed in a way that alters what counts as high risk, and whether your earlier passes still hold. The repository also ships KNOWN_LIMITATIONS.md and TEST_CASES.md, which are the files to read before assuming a new version behaves like the old one. Budget time for that reading, not just for the upgrade.

Editorial conclusion

Adopt it if you are revising a master's thesis chapter by chapter and want to keep your own wording, citations, variables and p-values while breaking repeated structures; the README's forbidden-operations list is the part to read before you let it touch anything. Do not adopt it if you want a guaranteed Turnitin AI or CNKI AIGC score, whole-chapter regeneration, or a third-party humanizer in disguise, because the project explicitly rules those out. Verify first that your workflow can supply an external detector report as a calibration signal and that you can run a comparable external re-test afterwards, since the skill treats scores as external evidence rather than something it can predict.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. Moonlit-Pages/AIGC-Detector-Rewriter-Skill on GitHub
4. README
5. Releases